0

我正在努力将 tensorflow 2.0 dcGAN 教程 ( https://www.tensorflow.org/beta/tutorials/generation/dcgan ) 中的代码改编为音频信号的频谱图。我正在使用 libroasa chroma_cqt 将原始音频数据转换为 WxHx2 矩阵，并将其用作输入。当我尝试通过放大随机噪声来创建种子矩阵时，我得到的结果是随机噪声和 0 的时空交替带和顶部的细黑条（见图）。裸露的噪音

我已经修改了原始教程代码以处理各种尺寸的图像，并为种子图像和最终输出提供了良好的结果，但是相同的原则并没有引导我使用 3 维数据。如何确保我正在制作具有适当覆盖率的种子，而不是在实际训练模型时继续问题？

from __future__ import absolute_import, division, print_function, unicode_literals

import tensorflow as tf

tf.__version__

import numpy as np
import os
from tensorflow.keras import layers
import librosa
import librosa.display

import matplotlib.pyplot as plt

os.environ['CUDA_VISIBLE_DEVICES'] = '-1'

sr = 44100/2
sample_path = os.getcwd()


def make_generator_model():
    model = tf.keras.Sequential()
    model.add(layers.Dense(2*7*19*128, use_bias=False, dtype='float32', input_shape=(361,)))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Reshape((2 ,7, 19, 128)))
    assert model.output_shape == (None,2, 7, 19, 128) # Note: None is the batch size

    model.add(layers.Conv3DTranspose(128, (2, 5, 5), strides=(1, 6, 1), padding='same', use_bias=False))
    assert model.output_shape == (None, 2, 42, 19, 128)
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Conv3DTranspose(128, (2, 5, 5), strides=(1, 3, 19), padding='same', use_bias=False))
    assert model.output_shape == (None, 2, 126, 361, 128)
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Conv3DTranspose(1, (2, 5, 5), strides=(1, 2, 1), padding='same', use_bias=False, activation='tanh'))
    assert model.output_shape == (None, 2, 252, 361, 1)

    return model


generator = make_generator_model()
noise = tf.random.normal([1, 361])
generated_audio = generator(noise, training=False)


D = []
for x in range(len(generated_audio[0][0])):
    this_line = []    
    for y in range(len(generated_audio[0][0][x])):
        this_line.append(np.complex(generated_audio[0][0][x][y],generated_audio[0][1][x][y]))
    D.append(this_line)
D = np.asarray(D)


librosa.display.specshow(librosa.amplitude_to_db(np.abs(D), ref=np.max),
                          sr=sr, x_axis='time', y_axis='cqt_note')
plt.axis('off')
plt.savefig(sample_path + '\\image_at_epoch_fuzz.png')
plt.show()


print(D.shape)

我正在输出音频噪声的视觉表示，它应该看起来像一个完全模糊的图像。相反，我得到了交替的噪音和大的黑色垂直条和顶部的细黑条。

问题最终是我需要遵循哪些规则来匹配生成器种子、内核大小和步幅？有人可以提供一个示例，说明如何以编程方式确保给定层数的步幅和内核大小不匹配？

4

2 回答 2

1

当您的步幅过大时，就会发生这种情况。尝试使用更大的Dense层和更小的步幅，或者更多的Conv3DTranspose层。就像是：

def make_generator_model():
    model = tf.keras.Sequential()
    model.add(layers.Dense(2*32*46*128, use_bias=False, dtype='float32', input_shape=(361,)))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Reshape((2, 32, 46, 128)))
    # assert model.output_shape == (None,2, 7, 19, 128) # Note: None is the batch size

    model.add(layers.Conv3DTranspose(128, (2, 3, 3), strides=(1, 2, 2), padding='same', use_bias=False))
    # assert model.output_shape == (None, 2, 42, 19, 128)
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Conv3DTranspose(128, (2, 3, 3), strides=(1, 2, 2), padding='same', use_bias=False))
    # assert model.output_shape == (None, 2, 126, 361, 128)
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Conv3DTranspose(1, (2, 3, 3), strides=(1, 2, 2), padding='same', use_bias=False, activation='tanh'))
    # assert model.output_shape == (None, 2, 252, 361, 1)
    model.add(layers.Lambda(lambda x: x[:, :, :252, :361, :]))

    return model
于 2019-06-23T12:29:39.247 回答
0

所以问题最终是关于卷积 kernel_size 和步幅之间的关系（为了更好地解释每个术语，请参见此处的 Conv3DTranspose 部分 https://keras.io/layers/convolutional/）。密集层一开始就很好。在原始代码中，以下 Conv3DTranspose 行 kernel_size 不涵盖高度方向 (5<6) 和宽度方向 (5<19) 的步幅

model.add(layers.Conv3DTranspose(128, (2, 5, 5), strides=(1, 6, 1), padding='same', use_bias=False))

model.add(layers.Conv3DTranspose(128, (2, 5, 5), strides=(1, 3, 19), padding='same', use_bias=False))

通过确保 kernel_size 的最小尺寸与所选步幅尺寸相匹配来解决该问题。这是固定的代码：

from __future__ import absolute_import, division, print_function, unicode_literals

import tensorflow as tf

tf.__version__

import numpy as np
import os
from tensorflow.keras import layers
import librosa
import librosa.display

import matplotlib.pyplot as plt

os.environ['CUDA_VISIBLE_DEVICES'] = '-1'

sr = 44100/2
sample_path = os.getcwd()


def make_generator_model():
    model = tf.keras.Sequential()
    model.add(layers.Dense(2*7*19*128, use_bias=False, dtype='float64', input_shape=(50,)))
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Reshape((2 ,7, 19, 128)))
    #assert model.output_shape == (None,2, 7, 9, 128) # Note: None is the batch size

    model.add(layers.Conv3DTranspose(128, (1, 6, 1), strides=(1, 6, 1), padding='same', use_bias=False))
    #assert model.output_shape == (None, 2, 42, 19, 128)
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReL())


    model.add(layers.Conv3DTranspose(128, (1, 3, 19), strides=(1, 3, 19), padding='same', use_bias=False))
    #assert model.output_shape == (None, 2, 126, 361, 128)
    model.add(layers.BatchNormalization())
    model.add(layers.LeakyReLU())

    model.add(layers.Conv3DTranspose(1, (1, 2, 1), strides=(1, 2, 1), padding='same', use_bias=False, activation='tanh'))
    #assert model.output_shape == (None, 2, 252, 361, 1)

    return model


generator = make_generator_model()
noise = tf.random.normal([1, 50])
generated_audio = generator(noise, training=False)


D = []
for x in range(len(generated_audio[0][0])):
    this_line = []    
    for y in range(len(generated_audio[0][0][x])):
        this_line.append(np.complex(generated_audio[0][0][x][y],generated_audio[0][1][x][y]))
    D.append(this_line)
D = np.asarray(D)


librosa.display.specshow(librosa.amplitude_to_db(np.abs(D), ref=np.max),
                          sr=sr, x_axis='time', y_axis='cqt_note')
plt.axis('off')
plt.savefig(sample_path + '\\image_at_epoch_fuzz.png')
plt.show()


print(D.shape)

结果：正确上采样的噪声图像

于 2019-07-09T01:50:43.270 回答